A working record of the software I conceive, build, and answer for.

Scholium connects the pieces these organizations have to run: enrollment, attendance, grades, records, progress reports, transcripts, billing, parent access, payments, imports, and the administrative workflows in between. Progress reports are AI-assisted through the Anthropic API.

Security got its own phase, tested live rather than only reviewed: session rotation, login lockout, CSRF rejection, and the authentication flows themselves. Pricing stayed simple, with two flat plans, Co-Op and Learning Center, billed monthly or yearly.

It ran in production for about four months and is offline for now. The highlight reel below walks through it.

Qorva covers landed-cost planning from FOB to DDP, shipment tracking, and duty estimates that include Section 301 and 232 additional duties. Around that sit HS-code reference and search, purchase orders, document workflows, reconciliation, reporting, accounting-oriented exports, and live FX rates from frankfurter.app.

Duty estimates are only as good as the tariff data behind them, so the platform keeps its own: a pipeline that tracks the U.S. Harmonized Tariff Schedule revision by revision. Two flat plans, Standard and Pro, billed monthly or yearly.

Qorva reached production and is offline for now. The highlight reel below shows it at work.

The Decision Suite is the flagship demonstration for a planned fixed-scope consulting practice at rosalindgash.com: financial decision systems for founder-led SaaS companies. It is built around Havenpoint Workflow, a SaaS company with 24 months of data.

Note Havenpoint Workflow is fictional, and all of its data is synthetic.

Four connected modules share that one dataset: Performance & Cash, Accounting & Assets, Workforce & Hiring, and Tax Readiness & Expansion, all under an executive overview, the Founder Decision Center. Scenario controls for churn, new-customer pricing, new-business strength, a hire, and UK market entry flow through the 12-month forecast wherever they are financially relevant. A hire changes burn and runway; churn moves ending MRR and cash.

Every formula lives in one calculation library, never inside components, backed by a metric dictionary and an assumptions register. Charts are accessible, with data-table alternatives, as part of WCAG 2.1 AA work. The suite also includes a SaaS Financial Visibility Scorecard and a fail-closed contact form (server action, honeypot). rosalindgash.com is offline for now.

Ops was built to be the backend for rosalindgash.com, the site for my planned consulting practice around the SaaS Financial Decision Suite. It is the practice’s CRM, sales pipeline, proposals, contracts, projects, activities, and recurring appointments.

Clients get a token-scoped, read-only portal for document exchange. Ops is internal, so there is no public link.

Studio keeps Preview and Program separate, with a deliberate TAKE between them, plus scenes, a run of show, preflight checks, and autosave with recovery. A four-channel mixer (mic, media, system, master) handles mix-minus. It records up to 4K, MKV first and then remuxed to MP4, with seven separate stereo tracks, and streams to YouTube at 1080p30 over RTMPS with an automatic safety recording and bounded reconnects.

A YouTube chat panel feeds a Q&A queue and on-screen comment cards, alongside lower-thirds and ticker graphics. Up to three remote guests join over WebRTC, through a green room and private join links.

Post-production runs on the device, with no cloud and no per-use fees: dialogue isolation; a transcript editor built on Whisper; an episode producer that drafts titles, descriptions, chapters, and show notes with inspectable citations, using a local LLM through llama.cpp; clip drafts with 9:16 framing and captions; and thumbnail concepts from a local image model. I review everything before it is exported.

Studio shares a Core library with RosalindCam and comes with a 22-page printable user guide. It is a personal build, and real-device acceptance of the guest features is still in progress.

RosalindCam records screen only, webcam only, or screen and webcam together, with microphone audio, to 1080p MP4. The webcam can sit side by side with the screen or float as a movable, resizable picture-in-picture that can be adjusted while recording.

I wanted to hide my room on camera without a green screen, so it does real-time background replacement (a colour or an image) or blur, using the RobustVideoMatting model on the GPU through ONNX Runtime and DirectML, with a CPU fallback and edge controls.

Recording is crash-safe: it writes MKV first, then remuxes to MP4, using AMD hardware H.264 with a software fallback. It ships as a portable, self-contained .exe, backed by a 99-check native smoke suite.

PingBuoy was my first major AI-assisted SaaS. I took it from product planning through architecture, implementation, security work, payment integration, deployment, and debugging.

PingBuoy is archived, but a live demo loaded with sample data is still running, so you can use it rather than read about it.